Research Summary: VeriWeave Govern: Evidence-Gated Deterministic Runtime Governance for Enterprise AI Agents

The increasing use of Artificial Intelligence (AI) agents in enterprise settings, involving tool interaction, infrastructure modification, and handling of protected data, necessitates robust governance mechanisms. VeriWeave Govern is introduced as a deterministic runtime governance layer designed to separate AI agent action generation from authorization. This system evaluates agent actions against defined policies, validates evidence, and routes consequential actions for human review, while maintaining an auditable state.

Why it matters

The secure and responsible deployment of AI agents is paramount for enterprises, especially as these agents gain capabilities to interact with critical systems and sensitive information. Implementing robust governance solutions like VeriWeave Govern is essential to mitigate operational, security, and compliance risks associated with autonomous AI actions, ensuring trust and control over AI agent operations.

Key insights

• Enterprise AI agents interact with tools, modify infrastructure, and process protected data, creating a critical need for action authorization separate from action generation.
• VeriWeave Govern offers a deterministic runtime governance layer that processes structured agent actions.
• The system evaluates actions against versioned policies and validates typed evidence.
• It employs a fixed precedence logic of 'deny > review > allow' for action processing.
• Consequential actions are routed to accountable human review processes.
• The system records a replayable, tamper-evident audit state.
• GovernBench, an evaluation framework, tested VeriWeave Govern across 30 independent seeds and 60,000 oracle-labelled cases, covering five enterprise domains, adversarial evidence, out-of-distribution actions, and temporal policy evolution.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.37457